Injecting Semantics for Explaining Link Predictions

Knowledge Graphs, despite their vast utlity, often suffer from incompleteness. This issue stems from their complex and automated construction process, the inherent difficulties in data collection, and the dynamic nature of the real-world phenomena they reflect. Addressing this issue, numerous Link Prediction models have been developed to infer missing facts. Most of the existing methods are based on Knowledge Graph Embeddings which are numeric representations of the entities and relations in a Knowledge Graph. These approaches are effective and scale on very large Knowledge Graphs. However, they are not capable of providing explanations for their outcomes. For this reasons, beyond prediction, explaining the outcomes has gained significant attention. Addressing this concern is Kelpie, an emerging and promising framework for explanation generation, which is gaining considerable attention due to its adaptability to any embedding model. Given a predicted link Kelpie extracts as explanation subsets of relevant facts, i.e. facts that have enabled the model to predict the link. Notably, Kelpie searches for explanations within the sub-graph composed of the facts related to the prediction to explain, considering both individual facts and combinations thereof. However, this thorough strategy of considering combinations of multiple facts leads to a significant challenge: an explosion in the number of candidate explanations. We want to enhance this strategy with respect to three goals, firstly we seek to decrease the number of candidate explanations, furthermore, we aim to maximize the performance in terms of explanation quality and lastly we strive to represent the explanations on different levels of granularity. To achieve these goals, we propose to inject semantics in the search strategy. We aim to compute an abstraction of the search space. Specifically, we aim at a new search space which is more compact and focused on more general patterns in order to support our research goals. In particular, we propose to extract the types of the sub-graph nodes and compute the quotient set – the set of equivalence classes - in order to merge similar nodes in a single node. Then, use such types to build the quotient graph which connects the equivalence classes with edges trying to preserve as much as possible the original structural properties. An edge in the quotient graph connects two equivalence classes of nodes in the original sub-graph. Hence we can map each edge in the quotient graph to a batch of facts in the original sub-graph. We execute the Kelpie search strategy using the quotient graph as input. Specifically, we enumerate all the edges in the quotient graph and subsequently all their combinations. This is equivalent to enumerating batches of facts in the original graph. The abstraction introduced by the quotient graph holds the potential to enhance the performance by directly focusing on general patterns. However, there is also a risk of losing valuable information and degrading performance. For this purpose we propose alternative quotient formulations with different (incremental) granularity. Finally a fact in the quotient graph leads to a dual perspective of explanation, in terms of equivalence classes or in terms of individual entities within. The three formulations and Kelpie, as a baseline, are experimentally compared to find the optimal trade-off between abstraction and quality.